Task paradigm

We carried out 3 fMRI experiments to investigate the mental state representations of decision uncertainty in metacognition and mentalizing (Fig 2A; S1 Fig). A total of 28 healthy participants took part in all of the 3 experiments (see Methods). In experiment 1, each participant judged the gross motion direction of random moving dots and rated his/her uncertainty about the preceding decision (Fig 2B). There were 4 different task difficulty levels randomly mixed in the task (S2 Fig). Hereafter, this perceptual decision-making task was referred to as the random dots motion (RDM) task (Fig 2A). The participant reported the current-self decision uncertainty (CS-DU), immediately accompanying perceptual decision-making in each trial.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Task paradigms and behavioral results.

(a) The fMRI experimental setup. During the metacognition task, the participant inside the MRI scanner did the metacognition task alone. During the mentalizing tasks, the participant inside the MRI scanner observed the task performance on the metacognition task done by the target participant who was outside the scanner. (b) The metacognition task: The participant completed the RDM task and reported his/her CS-DU. (c) The type 1 mentalizing tasks: The participant observed the RDM task performance by a target participant and reported the target participant’s decision inaccuracy. The target participant was either an AO-DI or the PS-DI. (d) The type 2 mentalizing task: Instead of judging the target participants’ decision inaccuracy, the participants estimated decision uncertainty that would be concurrently reported by the AO/PS in the current trial (AO-DU/PS-DU). (e) The decision inaccuracy changed with task difficulty and RT in the AO-DU task, averaged across participants (n = 28). (f) Theoretically, decision inaccuracy is a sigmoid function of task difficulty and RT on each trial. Each target participant has unique internal noise (σ₁) in perceptual decision-making that causes different variances in decision inaccuracy. (g) Theoretically, in estimating the target participant’s decision uncertainty, the unique internal noise (σ₂) needs to be further considered in mapping decision inaccuracy to decision uncertainty. (h) Theoretically, different levels of internal noise (σ₂) cause different metacognitive abilities (AUROC). (i) The regression beta values of the normalized (task) difficulty and RT with the estimates in each task. The weights did not differ across mentalizing tasks (ANOVA, task difficulty: F_[3,112] = 0.11, P = 0.95; RT: F_[3,112] = 0.16, P = 0.92), but significantly differ from the metacognition task (2-tailed paired t test, task difficulty: t₂₇ = 6.2; P = 2.5 × 10⁻⁸; RT: t₂₇ = 4.1; P = 5.9 × 10⁻⁵). (j) The correlation between the estimated decision inaccuracy/uncertainty in the mentalizing tasks with the target participant’s actual decision uncertainty reported in the metacognition task, before (original: ANOVA, F_[3,112] = 0.28, P = 0.84) and after (residual: Ps > 0.30) the associations with external cue information were regressed out. (k) The consistency between the estimated decision inaccuracy/uncertainty and the actual decision outcome (true or false) measured by the AUROC, before (original: ANOVA, F_[4,139] = 1.22, P = 0.31) and after (residual: Ps > 0.20) the associations with external cue information were regressed out. (l) The ratio of estimate residuals to the total estimate variances (2-tailed t test, AO: t₂₇ = 2.5; P = 0.0096 in the contrast between the type 2 and type 1 mentalizing tasks; PS: t₂₇ = 2.1; P = 0.023 in the contrast between the type 2 and type 1 mentalizing tasks; 2-tailed paired t test, t₂₇ = 3.5; P = 3.3 × 10⁻⁴ in the contrast between the metacognition task and the mentalizing tasks). The error bars represent SEM across participants. *P < 0.05; **P < 0.01; ***P < 0.001, after Bonferroni correction. The raw data for Fig 2 can be found in the Supporting information as S1 Data. ANOVA, analysis of variance; AO-DI, anonymous other decision inaccuracy; AUROC, area under the ROC curve; CS-DU, current-self decision uncertainty; fMRI, functional magnetic resonance imaging; PS-DI, past-self decision inaccuracy; RDM, random dots motion; RT, reaction time; SEM, standard error of the mean.

https://doi.org/10.1371/journal.pbio.3001301.g002

In experiment 2, the participant inside the scanner observed an anonymous other (AO) concurrently performing the RDM task outside the scanner and judged the AO’s decision inaccuracy (AO-DI). Decision inaccuracy is the opposite of decision accuracy where decision accuracy is the objective probability that the AO’s decision is correct (Fig 2C). Differing from the metacognition task (CS-DU), the partner’s cognitive states were inaccessible. It thus necessitated the participant to infer the probability that the partner’s decision was correct. To avoid evoking the participant’s own decision uncertainty, the stimuli presented to the participant were noiseless: Only coherently moving dots were moving, whereas randomly moving dots remained stationary. By virtue of this altered stimulus presentation, the participant could perceive the task difficulty without evoking his/her own decision uncertainty (Fig 2C). This is a necessary condition to dissociate the neural representations of decision uncertainty in mentalizing from those in metacognition. Otherwise, the participant might use his/her own decision uncertainty to estimate the partner’s decision uncertainty. The partner’s response time (RT) was reported to the participant by a progress color bar, whereas neither the choice nor the reported decision uncertainty by the partner was presented to the participant. Hence, the participant could only use the external cue information of task difficulty and RT to estimate the partner’s decision inaccuracy. In a parallel task, the participant instead observed task performance on the metacognition task done past by himself/herself and judged the past-self decision inaccuracy (PS-DI). Otherwise, the experimental procedure was identical to the AO-DI task. Notably, as the past decision-making processes by oneself were also inaccessible and the past mental states associated with similar stimuli were impossible to explicitly memorize, the underlying cognitive process might also be mentalizing. We therefore refer to the 2 tasks as the type 1 mentalizing tasks.

In experiment 3 (Fig 2D), the experimental procedure was identical to experiment 2, but the participant estimated the target participants’ mental states of decision uncertainty in each trial, namely the participant estimated the target participants’ believes about whether their own decisions were correct. The 2 tasks thus also entailed mentalizing to attribute type 2 mental states of decision uncertainty to the AO/PS (AO-DU and PS-DU). We therefore refer to the 2 tasks as the type 2 mentalizing tasks.

Experiment 1 and experiment 3 were conducted in the same session, but experiment 2 was conducted in another session. To reduce confusion between type 1 and type 2 mentalizing tasks, the 2 sessions were separated at least over 2 weeks. The tasks in each session were randomly interleaved and were counterbalanced across the participants.

The task sequences of the 4 mentalizing tasks were identical, only the instructions differed. Thereby, any behavioral and neural differences between them should be caused by different mentalizing processes. The task sequences of the metacognition task and the mentalizing tasks were also quite similar. However, the differences between the 2 types of tasks existed in both the perception phase and the judgment phase. Here, however, we are not so much concerned with the former but with the latter phase. In particular, the participant currently experienced decision uncertainty accompanying perceptual decision-making in the metacognition task, but inferred decision uncertainty that was not concurrently experienced from the cue information in the mentalizing tasks. However, because the perception phase and the judgment phase were temporally close to each other, it might be argued that any difference in neural signatures is due to the difference in the stimulus presentation as opposed to the difference between the metacognitive and metalizing processes. To confirm whether the neural correlates between the 2 phases are separable, we made analyses on the simulated fMRI signals generated by the same task sequence. Our simulation analyses demonstrated that the neural correlates of decision uncertainty in the perception phase or the judgment phase could be dissociated by conventional general linear models (GLMs) (S3A Fig).

Hierarchical mental state representations of decision inaccuracy and decision uncertainty in mentalizing

To assess behavioral metrics used for data analyses, we made theoretical analyses on mental state representations of decision inaccuracy and decision uncertainty in mentalizing. According to the decision-making theory [27], decision inaccuracy is crucially dependent on both task difficulty and RT. The higher the task difficulty and the longer the RT, the higher the decision inaccuracy (Fig 2E). For the sake of simplicity, decision inaccuracy is assumed to be a sigmoid function of task difficulty and RT (Eq 1 in Methods; Fig 2F). Hence, it is plausible to estimate decision inaccuracy and decision uncertainty in the mentalizing tasks merely from external information provided by task difficulty and RT. However, one indispensable process to distinguish social inferences in mentalizing from nonsocial inferences or associations is taking the target participant’s perspective. For example, in the type 1 mentalizing tasks, the participants should consider that the target participant has unique internal noise (σ₁) during the perceptual decision-making process as described by the drift-diffusion model [27], which affects the target participant’s object-level performance (i.e., decision inaccuracy, Fig 2F). In the type 2 mentalizing tasks, the participant should further consider that the target participant has unique internal noise (σ₂) in mapping decision inaccuracy to decision uncertainty (Fig 2G), which renders the target participant’s unique metacognitive ability even with the same object-level performance [28] (i.e., a low variance results in a high metacognitive ability). We constructed the receiver operating characteristic (ROC) curve by using the level of decision uncertainty as the criterion to judge the incorrectness of the choice in each trial and measured the metacognitive ability as the area under the ROC curve (AUROC), indicating the extent to which the subjective uncertainty ratings matched the actual decision inaccuracy [28] (Fig 2H). Taken together, the internal mental state representations of decision inaccuracy and decision uncertainty in mentalizing should be hierarchically organized. However, due to the lack of feedback in the mentalizing tasks, each participant did not learn about the target participants’ (even for the PS) object-level and meta-level performance. Therefore, the internal information generated by mentalizing may not reflect the target participant’s actual internal mental states.

Behavioral results

We analyzed the behavioral data in the experiments to assess how the participants used the associations with external cue information of task difficulty and RT to estimate the targets’ decision inaccuracy and decision uncertainty. The analyses showed that the weights of normalized task difficulty and RT (Eq 1 in Methods) on the estimates were equivalent [analysis of variance (ANOVA), task difficulty: F_[3,112] = 0.11, P = 0.95; RT: F_[3,112] = 0.16, P = 0.92; Fig 2I] and were highly correlated across the mentalizing tasks (S4 Fig). Hence, the participants used such external cue information to estimate the corresponding mental states equally across the mentalizing tasks. Notably, the estimates in the mentalizing tasks relied more on task difficulty than RT (2-tailed paired t test, t₂₇ = 6.2; P = 2.5 × 10⁻⁸). On the contrary, the estimates of decision uncertainty in the metacognition task relied more on RT than task difficulty (2-tailed paired t test, t₂₇ = 4.1; P = 5.9 × 10⁻⁵; Fig 2I). This is likely due to the fact that stimulus coherence (by virtue of the nature of the experimental design) was clearly discerned in the mentalizing tasks, but it was hard to inversely infer the stimulus coherences in the metacognition task. Because of the stable associations with external cue information, the decision inaccuracy/uncertainty estimated by the participants in the mentalizing tasks was correlated with what the target participants had actually reported themselves at each level of difficulty and RT in the metacognition task (ANOVA, F_[3,112] = 0.28, P = 0.84, Fig 2J).

However, all these correlations in the mentalizing tasks disappeared after regressing out the associations with external cue information of task difficulty and RT from the estimates of decision inaccuracy/uncertainty reported by the participants (2-tailed t test, Ps > 0.30; Fig 2J). That is, estimate residuals in each mentalizing task did not further predict the actual decision uncertainty reported by the target participants in the metacognition task. Further, as the estimates of decision uncertainty in each task could largely predict the actual decision inaccuracy, we used the AUROC to characterize this consistency. On average, the AUROCs (mean: 0.67 to 0.71) were larger than the chance level that was calculated by shuffling the orders between the estimates and actual decision inaccuracy and showed no significant differences across all the tasks (ANOVA, F_[4,139] = 1.22, P = 0.31, Fig 2K). After the associations with external cue information were regressed out, the residual AUROCs (measured by estimate residuals) were no longer significantly different from the chance level in each mentalizing task (2-tailed t test, Ps > 0.20), but it remained significant in the metacognition task (2-tailed t test, t₂₇ = 11.8; P = 3.6 × 10⁻¹²; Fig 2K). Estimate residuals in the metacognition task should reflect the subjective difficulty due to the trial-by-trial noises in internal neural processing that task difficulty (coherence) and RT could not explain [29]. Thereby, the residuals also considerably contributed to decision uncertainty in the metacognition task. In striking contrast, reliable estimates of decision inaccuracy/uncertainty in the mentalizing tasks were crucially dependent on external information provided by task difficulty and RT.

Nonetheless, estimate residuals accounted for about half of the total variance of the estimates of decision inaccuracy/uncertainty in each mentalizing task (Fig 2L), although each ratio was much lower than that in the metacognition task (2-tailed paired t test, t₂₇ = 3.5; P = 3.3 × 10⁻⁴; Fig 2L). These estimate residuals might serve as a proxy of internal mental states that were generated through target participant’s perspective taking and were independent of external cue information. We thus used estimate residuals as the main behavioral metric to identify the neural representations of internal mental states. As indirect evidence, the variances of estimate residuals were significantly larger in the type 2 mentalizing tasks than the type 1 mentalizing tasks (2-tailed t test, AO: t₂₇ = 2.5; P = 0.0096; PS: t₂₇ = 2.1; P = 0.023). These extra variances might be generated by the additional process of the target participant’s meta-level perspective taking in the type 2 mentalizing tasks, as suggested by the hierarchical mental state representation model described above (Fig 2G and 2H).

Common neural representations of external cue information in mentalizing

We first examined the neural representations of external cue information in mentalizing by analyzing the fMRI data acquired during the experiments. As external cue information of task difficulty and RT contributed to the estimates equally across the mentalizing tasks, we hypothesized that the neural representations of each type of external cue information might be shared across the mentalizing tasks. To test the hypothesis, we regressed the trial-by-trial fMRI activities during the judgment phase with external cue information of task difficulty and RT across the whole brain in each task (see Methods). Across the mentalizing tasks, the fMRI activities in the primary visual cortex (V1) were negatively correlated with the levels of task difficulty (the conjunction analysis, z > 2.6, P < 0.05 after cluster-level family-wise error (FWE) correction; Fig 3A), decreasing as the number of moving dots was reduced (task difficulty increased). On the other hand, the fMRI activities in the right inferior parietal lobe (IPL) were positively correlated with the levels of task difficulty (the conjunction analysis, z > 2.6, P < 0.05 after cluster-level FWE correction; Fig 3A). In contrast, the fMRI activities in a wide range of brain regions were positively correlated with the levels of RTs (the conjunction analysis, z > 2.6, P < 0.05 after cluster-level FWE correction; Fig 3B). Among these brain regions, the right IPL region overlapped with the regions associated with task difficulty: The same right IPL region responded to both task difficulty and RT across the mentalizing tasks (Fig 3C and 3D). Thus, integration of the 2 pieces of external information together in the right IPL partially contributed to the estimates of decision inaccuracy/uncertainty in mentalizing.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Common neural representations of external cue information across the mentalizing tasks.

(a) The activation maps for the activities significantly correlated with task difficulty in a conjunction analysis across the mentalizing tasks (z > 2.6, P < 0.05 after cluster-level FWE correction). (b) The activation maps for the activities significantly correlated with RT in a conjunction analysis across the mentalizing tasks (z > 2.6, P < 0.05 after cluster-level FWE correction). The activation maps in (a) and (b) were displayed in radiological convention (the left/right side of the image corresponds to the right/left side of the brain). (c) The parametric regression beta values of task difficulty in each mentalizing task in the right IPL ROI defined by the conjunction of (a) and (b). (d) The beta values of RT in each mentalizing task in the right IPL ROI defined by the conjunction analysis of (a) and (b). The error bars represent SEM across participants. ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001, uncorrected. The raw data for Fig 3 can be found in the Supporting information as S1 Data. AO-DI, anonymous other decision inaccuracy; AO-DU, anonymous other decision uncertainty; FWE, family-wise error; IPL, inferior parietal lobe; PS-DI, past-self decision inaccuracy; PS-DU, past-self decision uncertainty; ROI, region of interest; RT, response time; SEM, standard error of the mean.

https://doi.org/10.1371/journal.pbio.3001301.g003

Distinct neural representations of estimate residuals in mentalizing

We then examined the neural representations of internal mental states that were unassociated with external cue information in mentalizing. According to our theoretical analyses as described above, additional unique processes of target participant’s perspective taking should be involved in each mentalizing task besides the associations with external cue information. To explore the neural signatures, we regressed estimate residuals with the trial-by-trial fMRI activities during the judgment phase across the whole-brain voxels in each task. These neural correlates thus illustrated internal mental state representations that were unassociated with external information. The modulation effects were predominately in the judgment phase by comparison to the alternative GLM accounting for the modulation effects in the perception phase (S3B Fig).

When the participant estimated the AO’s decision uncertainty, estimate residuals were significantly correlated with the fMRI activities in the dorsomedial prefrontal cortex (dmPFC; red in Fig 4A), dorsally and anteriorly neighboring, but separate from, the dorsal anterior cingulate cortex (dACC) region representing the CS-DU in metacognition (blue in Fig 4A; see below), as well as in the left temporoparietal junction (TPJ) and the left inferior frontal junction (IFJ) (z > 3.1, P < 0.05 after cluster-level FWE correction, Fig 4A; see also S1 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Distinct neural representations of estimate residuals in metacognition and mentalizing.

(a) Neural correlates of estimate residuals of decision uncertainty in the metacognition task (dACC and lFPC, blue) and neural correlates of estimate residuals of decision uncertainty in the AO-DU (type 2) mentalizing task (dmPFC, left IFJ and left TPJ, red). (b) Neural correlates of estimate residuals of decision inaccuracy in the AO-DI (type 1) mentalizing task (left IFJ and left TPJ, green). (c) Neural correlates of estimate residuals of decision uncertainty in the PS-DU (type 2) mentalizing task (left lFPC, magenta). (d) Neural correlates of estimate residuals of decision inaccuracy in the PS-DI (type 1) mentalizing task (right-IFJ and precuneus, cyan). All the activation maps were displayed in radiological convention. (e) The comparisons of the parametric regression beta values with estimate residuals across tasks in the ROIs of dACC, dmPFC, left IFJ, and left TPJ (see also S5 Fig). The error bars represent SEM across participants. ns, not significant; *P < 0.05; **P < 0.01 after Bonferroni correction. The raw data for Fig 4 can be found in the Supporting information as S1 Data. AO-DI, anonymous other decision inaccuracy; AO-DU, anonymous other decision uncertainty; CS-DU, current-self decision uncertainty; dACC, dorsal anterior cingulate cortex; dmPFC, dorsomedial prefrontal cortex; IFJ, inferior frontal junction; lFPC, lateral frontopolar cortex; PS-DI, past-self decision inaccuracy; PS-DU, past-self decision uncertainty; ROI, region of interest; SEM, standard error of the mean; TPJ, temporoparietal junction.

https://doi.org/10.1371/journal.pbio.3001301.g004

When the participant estimated the AO’s decision inaccuracy, estimate residuals were also correlated with the fMRI activities in the left TPJ and the left IFJ (z > 3.1, P < 0.05 after cluster-level FWE correction, Fig 4B and 4E; see also S1 Table), but not in the dmPFC (2-tailed t test, t₂₇ = 1.1, P = 0.12; Fig 4E).

The dmPFC was selectively involved in type 2 mentalizing (Fig 4E; the post hoc comparisons with the other tasks: P < 0.0037 after Bonferroni correction), but the TPJ was involved in both type 1 and type 2 mentalizing (Fig 4E; 2-tailed paired t test, t₂₇ = 0.67, P = 0.21 between the 2 mentalizing tasks, the post hoc comparisons with the other tasks: P < 0.042 after Bonferroni correction). To further test the reliability of the dmPFC selectivity in type 2 mentalizing, we repeated the same GLM analysis on the dmPFC and TPJ regions independently defined by meta-analytical maps from the NeuroSynth database [30], as well as the conjunction regions between the meta-analytical regions and those in the current study. In both analyses, the results consistently support that the dmPFC but not the TPJ showed neural activities selective to type 2 mentalizing (S5 Fig).

Instead, estimate residuals of the past-self decision uncertainty (PS-DU) were selectively correlated with the fMRI activities in the left lateral frontopolar cortex (lFPC) (z > 3.1, P < 0.05 after cluster-level FWE correction, Fig 4C; but not significantly different from the other tasks, S6F Fig), whereas estimate residuals of the PS-DI were correlated with the fMRI activities in the precuneus (z > 3.1, P < 0.05 after cluster-level FWE correction, Fig 4D; the comparisons with the other tasks: P < 0.043 after Bonferroni correction, S6E Fig). Notably, both the left lFPC and precuneus regions were also shared with the metacognition-associated areas (S6E and S6F Fig).

In accordance with our theoretical account, the existence of different neural correlates of estimate residuals in the mentalizing tasks suggested that there are diverse neural representations of internal mental states during mentalizing in different social contexts. Notably, the right IPL that encoded task difficulty and RT did not represent estimate residuals in any of the mentalizing tasks (S6A Fig). The neural representations of internal mental states suggest that internal mental state representations do separately coexist with external cue associations during mentalizing (attributing covert mental states to the other target participants) at least, in humans. This was, however, could not have been inferred from behavioral observation alone.

Neural representations of estimate residuals in metacognition

In the metacognition task, estimate residuals were significantly correlated with the fMRI activities in the dACC and the lFPC (z > 3.1, P < 0.05 after cluster-level FWE correction, blue in Fig 4A; see also S1 Table), as repeatedly observed in previous studies [15–17,31,32]. Although the lFPC region was also associated with type 2 mentalizing (S6F Fig), the dACC region selectively represented estimate residuals in the metacognition task, but not in the mentalizing tasks (S6B Fig; the post hoc comparisons between the 2 types of tasks, P < 0.0031 after Bonferroni correction). Notably, the components of decision uncertainty associated with task difficulty and RT were also represented in the dACC (S6B Fig). Thus, the dACC uniformly represented all components of decision uncertainty in metacognition.

Common neural representations of estimate uncertainty across the mentalizing tasks

In the mentalizing tasks, the use of external cue associations cannot provide sufficient information to estimate the target participants’ trial-by-trial decision inaccuracy/uncertainty, which should be also much affected by random neural processes characterized by each target participant’s unique internal noises (σ₁ and σ₂, Fig 2F and 2H). Thereby, the estimating processes in mentalizing are often accompanied by uncertainty about the estimates, namely estimate uncertainty [33,34]. Hence, the estimate theory predicts that estimate uncertainty should be higher when the estimates are at the middle levels as opposed to the lowest or the highest level (inverted U-shape). The levels of estimate uncertainty could be appreciated by the RTs used for reporting the estimates. Longer RTs indicate higher uncertainty in making estimates. The RTs were longer at the middle levels than at the lowest or highest level (inverted U-shape, quadratic regression, 2-tailed t test, Ps < 0.001; Fig 5A). Further, we divided all trials equally into 8 bins according to the quantities of external information that were calculated by a sigmoid function of task difficulty and RT after fitting with the estimated decision inaccuracy/uncertainty (Eq 1 in Methods). The RTs in reporting the estimates were also longer at the middle bins than at the lower and higher bins (inverted U-shape, quadratic regression, 2-tailed t test, Ps < 0.001; Fig 5B). For the sake of simplicity, we operationally defined the trial-by-trial estimate uncertainty as the negative value of the deviation from the mean of the estimates in each trial, i.e.,–|estimate-mean(estimates)|. By virtue of this definition, estimate uncertainty was higher at the middle levels of the estimates (Fig 5C) and the quantities of external information (Fig 5D). Notably, estimate uncertainty was statistically uncorrelated with the estimates per se in each task.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Common neural representations of estimate uncertainty across the mentalizing tasks.

(a) The inverted U-shape relationships between RTs for reporting the estimates and the estimates commonly in the mentalizing tasks; (b) The inverted U-shape relationships between RTs for reporting the estimates and external information calculated by a sigmoid function of task difficulty and RT. (c) estimate uncertainty defined as (–1*|estimate–mean(estimates)|) had an inverted U-shape relationship with the estimates. (d) Estimate uncertainty had an inverted U-shape relationship with external information. (e) The neural activities were significantly correlated with estimate uncertainty in each mentalizing task. (f) The dACC was commonly correlated with estimate uncertainty (the conjunction analysis, yellow), overlapping with those correlated with decision uncertainty in the metacognition task (blue, see also Fig 4A). All the activation maps were displayed in radiological convention. (g) The inverted U-shape relationships between residual variances and external information. (h) The dACC activities had an inverted U-shape relationship with external information. (i) The linear regression beta values of the dACC activities with the residual variances. (j) The residual magnitudes were larger at the middle levels of external information. (k) The dACC activities were larger when the residual magnitudes were larger. (l) The linear regression beta values of the dACC activities with residual magnitudes. (m) The trials were divided into 3 subgroups with the same residual variances but different residual magnitudes (HV-HM: high_variance-high_magnitude versus HV_LM: high_variance-low_magnitude) and the same residual magnitudes but different residual variances (HV-LM versus LV_LM: low_variance-low_magnitude). (n) The dACC activities were significantly different between different residual variances, but not between different residual magnitudes. The error bars represent SEM across the participants. ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001, uncorrected. The raw data for Fig 5 can be found in the Supporting information as S1 Data. AO-DI, anonymous other decision inaccuracy; AO-DU, anonymous other decision uncertainty; CS-DU, current-self decision uncertainty; dACC, dorsal anterior cingulate cortex; PS-DI, past-self decision inaccuracy; PS-DU, past-self decision uncertainty; RT, response time; SEM, standard error of the mean.

https://doi.org/10.1371/journal.pbio.3001301.g005

We then examined neural correlates of estimate uncertainty in each task. The neural correlates of estimate uncertainty in each mentalizing task were similar to those of decision uncertainty in metacognition (S7 Fig). Commonly across the mentalizing tasks, the fMRI activities in the dACC were significantly and positively correlated with estimate uncertainty (Fig 5E). Critically, the dACC region associated with estimate uncertainty across the mentalizing tasks was largely overlapping with the dACC region associated with decision uncertainty in the metacognition task (Fig 5F). Notably, this is complementary with prior findings that the fMRI activities in the ventromedial prefrontal cortex (vmPFC) and the posterior cingulate cortex (PCC) are responsive to confidence about the estimates of different properties of the presented stimuli [33].

dACC was involved in monitoring the mentalizing processes

According to the estimation theory [33,34], estimate uncertainty is derived from the variance of estimate residuals (residual variance), which is larger in the middle range of the quantities of external information (Fig 2F). Consisting with this prediction, the residual variance in each bin of the quantities of external information was a negative parabolic function in each task (inverted U-shape, quadratic regression, 2-tailed t test, Ps < 0.01; Fig 5G). Accordingly, the dACC activity averaged in each bin of the quantities of external information was also a negative parabolic function in each mentalizing task (inverted U-shape, quadratic regression, 2-tailed t test, Ps < 0.05; Fig 5H), but not in the metacognition task (quadratic regression, 2-tailed t test, t₂₇ = −1.9; P = 0.07). In each mentalizing task, the dACC activities increased as the residual variance increased (linear regression, 2-tailed t test, Ps < 0.05, Fig 5I). Hence, the dACC activities tracked residual variances, suggesting its potential role in monitoring the mentalizing processes.

However, larger variances of estimate residuals might be accompanied by larger magnitudes of estimate residuals (residual magnitudes). To examine their concurrences, we calculated the mean residual magnitude averaged in each bin of the quantities of external information. Residual magnitudes had a similar relationship with the bins of external information as residual variances did (quadratic regression, 2-tailed t test, Ps < 0.001; Fig 5J). Accordingly, the dACC activity in each bin according to the signed values of estimate residuals (the mean was zero) had a positive parabolic function in each of the mentalizing tasks, as well as in the metacognition task (U-shape, quadratic regression, 2-tailed t test, Ps < 0.001; Fig 5K). Thus, the dACC activities also tracked residual magnitudes (linear regression, 2-tailed t test, Ps < 0.001, Fig 5L).

As the dACC has been suggested to play a critical role in adaptive control [35–37], it remains possible for an alternative interpretation that the dACC might play a control role in adaptively generating estimate residuals. That is, the larger the dACC activities were, the larger residuals were generated. To distinguish the 2 potentially alternative functional roles of the dACC involvement in mentalizing, we further conducted partial control analyses. We segregated all trials into 3 categories according to residual variances and residual magnitudes. To do so, we first median split all trials according to residual variances (Fig 5G, low variances: blue zone, high variances: red zone). Accordingly, the residual magnitudes should be also low in the former group (LV_LM: low variances and low magnitudes). Further, the residual magnitudes in the latter group were segregated into low (HV_LM: high variances and low magnitudes) and high (HV_HM: high variances and high magnitudes) subgroups by another median split (Fig 5M). Thereby, the residual variances were different between LV_LM and HV_LM (2-tailed paired t test, Ps < 6.7 × 10⁻¹⁴; Fig 5M), while their magnitudes were not different (2-tailed paired t test, Ps > 0.05; Fig 5M). In contrast, the residual magnitudes were different between HV_LM and HV_HM (2-tailed paired t test, Ps < 7.8 × 10⁻¹³; Fig 5M), while their variances were not different (2-tailed paired t test, Ps > 0.18; Fig 5M). We then could test the 2 alternative hypotheses by examining whether the dACC activities were selectively responsive to residual variances (involvement in monitoring) or residual magnitudes (involvement in generating). If the dACC activities were sensitive to residual variances, but not residual magnitudes, then the dACC activities should be significantly different between LV_LM and HV_LM, but not between HV_LM and HV_HM. Otherwise, the dACC activity pattern should be reverse. Our results support the former prediction (Fig 5N). That is, the dACC should be involved in monitoring the process of generating internal mental states, rather than directly in the generating process. Hence, metacognition monitors the mentalizing processes.

Although estimate uncertainty and decision uncertainty have been suggested to be distinct [34], our results illustrate that the neural representations of decision uncertainty in metacognition (the inner layer in Fig 6) and estimate uncertainty in mentalizing (the outer layer in Fig 6) were convergingly registered in the dACC. On the contrary, mentalizing was functionally irrelevant to metacognition and was neurally independent of the dACC activities. Instead, different forms of mentalizing had diverse formats of internal mental state representations in the human brain (the middle layer in Fig 6). Mental state representations of decision uncertainty in metacognition and mentalizing form a nested structure (Fig 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. The nested structure of neural representations of decision uncertainty in metacognition and mentalizing.

(a) The inner layer illustrates the neural representations of decision uncertainty in the dACC during monitoring one’s own decision-making by metacognition. (b) The middle layer illustrates distinct neural representations of decision uncertainty during monitoring the different target participants’ meta-level and object-level performance (left: the PS; right: others). The TPJ represents others’ decision inaccuracy (mental construct of the object-level performance) and the dmPFC represents others’ decision uncertainty (mental construct of the metacognitive performance). In contrast, the precuneus represents the PS-DI, and the lFPC represents the PS-DU. (c) The outer layer illustrates neural representations of estimate uncertainty in the dACC during the different mentalizing tasks, denoting that metacognition monitors mentalizing. dACC, dorsal anterior cingulate cortex; FPC, frontopolar cortex; dmPFC, dorsomedial prefrontal cortex; lFPC, lateral frontopolar cortex; PS, past self; PS-DI, past-self decision inaccuracy; PS-DU, past-self decision uncertainty; TPJ, temporoparietal junction.

https://doi.org/10.1371/journal.pbio.3001301.g006

Participants

We recruited 28 healthy right-handed participants (22 females, age: 23.5 ± 1.5 years old) to take part in all 5 tasks across 3 fMRI experiments. Informed consent was obtained from each participant in accordance with a protocol approved by Beijing Normal University Research Ethics Committee (ICBIR_A_0091_004).

Experiments

There were 3 fMRI experiments in the current study. In experiment 1, the participants carried out the metacognition task alone. In experiment 2, the participants inside the scanner observed the target participants (an AO or the PS) performing the metacognition task outside the scanner and estimated the target participants’ object-level performance, i.e., decision inaccuracy (AO-DI and PS-DI). The 2 tasks were then referred to as the type 1 mentalizing tasks. In experiment 3, the participants inside the scanner instead estimated the target participants’ meta-level performance, i.e., decision uncertainty (AO-DU and PS-DU). The 2 tasks were then referred to as the type 2 mentalizing tasks. Experiment 1 and experiment 3 were conducted in the same session, but experiment 2 was conducted in another session. The tasks in each session were randomly interleaved and were counterbalanced across the participants.

Metacognition task

In an aperture with a radius of 3 degrees of visual angle, 300 white dots (radius: 0.08 degrees, density: 2.0%) were randomly displayed on a black background (RDM). The dots moved in different directions at a speed of 8.0 degrees/second. The movement of each dot lasted 3 frames. A subset of dots moved coherently in the same direction (left, up, right, or down), while the other dots moved in different random directions. The participant was required to discriminate the net motion direction and rate his/her uncertainty about the decision by pressing a corresponding button. The uncertainty ratings were the reversals of confidence ratings. The participants were instructed to rate decision uncertainty at 4 levels (1 to 4). Level 1 indicated a belief of high probability that the decision was correct. Level 4 indicated a belief of low probability that the decision was correct (see S1 Fig). That is, the decision was believed close to the chance level. The task difficulty was determined by the percentage of coherently moving dots, with 4 levels of task difficulty randomly mixed in the task. The easiest and hardest levels were fixed at a coherence of 1% and 30%, respectively, while the 2 middle levels were set for each participant to achieve accuracy of 50% and 80%, determined by a staircase procedure in a practice session conducted several days prior to the fMRI experiments [15,31].

Mentalizing tasks

A pair of participants who had similar stimulus coherences at the 50% and 80% accuracy level in the practice session jointly took part in the task (the stimulus coherences at the 2 task difficulty levels used in the mentalizing tasks were the means of their stimulus coherences, respectively). One (the target participant) performed the metacognition task outside the scanner, and another (the observer participant) observed the target participant’s performance inside the scanner. The 2 computers separately presented the stimuli but were connected and synchronized by the local network following the TCP/IP protocol. Hence, they simultaneously performed their own tasks with the same sequence (see below) but in response to different stimuli and task requirements. To avoid eliciting the observer participant’s own decision uncertainty, the stimulus presented to her/him inside the scanner was noiseless: The randomly moving dots remained stationary, and only the coherently moving dots moved. While the target participant was responding, the observer participant saw a progress bar representing the elapsed time. However, the observer participant could not see the target participant’s choice and reported decision uncertainty. The observer participant then reported the estimate of the target participant’s decision inaccuracy/uncertainty. In the PS-DU and PS-DI mentalizing tasks, the target participant was replaced by the PS. That is, the observer participant observed RDM task performance previously done by himself/herself.

Task sequence

The sequences of the metacognition and mentalizing tasks were almost identical. In the metacognition task, each trial started with a green cross cue to indicate that the task stimulus would be presented 1 second later. The stimulus was then presented for 2 seconds, and 4 options of the moving directions were presented. The participant made a choice within 2 seconds. After a choice was made, 4 ratings from 1 (most certain) to 4 (most uncertain) were presented, and the participant reported the rating by pressing the corresponding button within 2 seconds. The intertrial interval (ITI) was jittered uniformly between 2 seconds and 6 seconds. On average, each trial lasted for 9 seconds. During the mentalizing tasks, although the sequence was identical to the metacognition task performed outside the scanner, the participant watched the noiseless stimulus presentation and the progress bar indicating elapsed time for the target participant’s responses and then reported the estimate rating of the target participant’s decision inaccuracy/uncertainty. Each task was conducted for 2 consecutive runs, each consisting of sixty trials.

Behavioral analyses

The reported ratings of decision inaccuracy/uncertainty in each of the metacognition and mentalizing tasks were associated with task difficulty and RT. We used a sigmoid function to characterize the relationship as follows: (1) where e_i represents the estimate of decision inaccuracy/uncertainty at the trial i, D_i and RT_i represent the task difficulty and the RT at the trial i, respectively, β₁ and β₂ represent the corresponding weights, and β₀ represents the constant bias, ε_i represents the estimate residual. Task difficulties and RT were separately normalized in each task for each participant. Further, for the mentalizing tasks, we also considered an autoregressive (AR) model to account for the associations of the estimates between the contiguous trials. However, the autocorrelation of the estimates between the contiguous trials in each mentalizing task was trivial, indicating that the estimates were not dependent on the history, but only on the current trial.

fMRI parameters

All fMRI experiments were conducted using a 3-T Siemens Trio MRI system with a 12-channel head coil (Siemens, Germany). Functional images were acquired with a single-shot gradient echo T₂* echo-planar imaging (EPI) sequence with volume repetition time of 2 seconds, echo time of 30 ms, slice thickness of 3.0 mm and in-plane resolution of 3.0 × 3.0 mm² (field of view: 192 × 192 mm²; flip angle: 90 degrees). A total of 38 axial slices were taken with interleaved acquisition, parallel to the anterior commissure-posterior commissure line.

fMRI analyses

fMRI analyses were conducted using FMRIB Software Library (FSL) [56]. To correct for rigid head motion, all EPI images were realigned to the first volume of the first scan. Data sets in which translation motion was larger than 2.0 mm or rotation motion was larger than 1.0 degree were discarded. No data were discarded in these analyses. Brain matter was separated from nonbrain using a mesh deformation approach and used to transform the EPI images to individual high-resolution structural images and then to the Montreal Neurological Institute (MNI) space by using affine registration with 12 degrees of freedom and resampling the data with a resolution of 2 × 2 × 2 mm³. Spatial smoothing with a 4-mm Gaussian kernel (full width at half-maximum) and high-pass temporal filtering with a cutoff of 0.005 Hz were applied to all fMRI data.

We used GLM to analyze the fMRI data. For the first-level analyses, 2 events were modeled in each trial. The first event (the perception phase) represented the stimulus presentation, time locked to the onset of the stimulus presentation, with a duration of the presentation time (2 seconds). The second event (the judgment phase) represented estimating decision inaccuracy/uncertainty, time locked to the onset of the rating, with the duration of the reaction time. The parametric modulation effects of task difficulty and RT, estimate residual, and estimate uncertainty were simultaneously added in the latter event (the judgment phase). Estimate residual was the residual of the estimate after the components associated with task difficulty and RT were regressed out. All the regressors were convolved with the canonical hemodynamic response function with double-gamma kennels. The results when the parametric modulation effects were added in the perception phase are presented in S3B Fig.

For the group-level analyses, we used FMRIB’s local analysis of mixed effects (FLAME), which models both “fixed effects” of within-participant variance and “random effects” of between-participant variance using Gaussian random-field theory in each task. Statistical parametric maps were generated with a threshold of z > 3.1, P < 0.05 after cluster-level FWE correction for multiple comparisons for each contrast, unless mentioned otherwise. The activation maps were displayed in radiological convention.

Conjunction analyses

We conducted conjunction analyses to test the common neural correlates of external information (task difficulty and RT) across the 4 mentalizing tasks, as well as the common neural correlates of the estimate uncertainty across the 4 mentalizing tasks. We identified significant regions in which there was evidence of effects in all contrasts of conditions using the FSL script easythresh. Statistical parametric maps were generated by the threshold with z > 2.6, P < 0.05 after cluster-level FWE correction for each contrast.

Region of interest (ROI) analyses

The IPL ROI was defined by the conjunction analysis across the 4 mentalizing tasks. The dACC ROI was defined in the CS-DU task. The dmPFC ROI was defined in the AO-DU task. The TPJ ROI was defined in the conjunction analysis between the AO-DU and AO-DI tasks. The FPC ROI was defined in the PS-DU task. The precuneus ROI was defined in the PS-DI task. To circumvent circular inference, we randomly analyzed the data from three-fourths of the participants to define each of the ROIs with a significance of z > 2.6, P < 0.05 after cluster-level FWE correction and then used the data from the held-out participants to obtain the parametric regression beta value in each ROI. We repeated this analysis one hundred times and averaged the parametric regression beta values for each ROI. For the dmPFC and TPJ ROIs, we further assessed meta-analytically derived ROIs associated with from Neurosynth (search term: “mentalizing”) [27], as well as the overlapping regions between our results and the Neurosynth ROIs. For the dACC region used for analyses in Fig 5, we used the overlap between the metacognition task and the mentalizing tasks obtained by conjunction analysis.

To identify whether the dACC activities in mentalizing tasks were associated with residual variances (i.e., in monitoring the mentalizing process) or residual magnitudes (i.e., in control of the mentalizing process), we sequentially divided the trials in each of the metacognition and mentalizing tasks into 3 subgroups according to the residual variances and the residual magnitudes. First, we divided all the trials equally into 8 bins according to the quantities of external information that were calculated by a sigmoid function of task difficulty and RT. We then split the bins of trials into 2 subgroups with low and high residual variances according to whether the residual variance in each bin was above the median and further split the trials with high residual variances into 2 subgroups with low and high residual magnitudes according to whether the residual magnitude of each trial was above the median (Fig 5J). We then calculated and compared the mean dACC activities among the 3 subgroups of trials (Fig 5K).

To test for the possibility of U-shaped relationships between dACC activity and estimate uncertainty, RT, or external information (Fig 5), we used a quadratic regression model (2) to fit with the data in each participant (polyfit in matlab) and tested the quadratic regression beta value (a) across the participants in comparison to zero using 2-tailed t test.

Simulations and parameter recovery

Because the stimulus presentation and uncertainty rating events were temporally close to each other, the GLM estimates for the 2 events should be highly collinear. To examine whether the GLM analyses could separate the underlying neural activity specifically associated with each of the 2 events, we generated fMRI data using the different parameters of the GLM as follows and validated the parameter recovery [29]. (3) where X_sti and X_rating represent the design matrix for the perception phase and the judgment phase, respectively. β₁ and β₃ are the mean activities of the 2 events, and β₂ and β₄ are the estimate uncertainty modulation effects on the 2 events, respectively. hrf is the canonical hemodynamic response function with 2-gamma kennels, and ℇ is an additional Gaussian noise. The values of β₁ and β₃, as well as β₂ or β₄, were independently and randomly drawn from a uniform distribution in the range of 0.2 to 0.8, while the alternative β₂ or β₄ was set to zero. That is, the modulation effect appeared only in either of the 2 phases, but not simultaneously in both phases. The signal-to-noise ratio (SNR) of the fMRI data were set uniformly in the range of 0.01 to 1 (the event-evoked fMRI signals are usually in the range of 0.1 to 10 of the noises). We then used the GLM with the mean activity and the parametric modulation effects in both phases to reconstruct the mean activity and the parametric modulation effects in the 2 events from the generated fMRI data with the sampling rate of 0.5 Hz (TR = 2 seconds). For each set of the parameters, 1,000 times of procedures were repeated, and the estimated values were averaged at each SNR level (S3 Fig).